If you want to do something that has approx. 1000 times more relevance than improving CPU schedulers by some statistically insignificant amount, then please solve the following practical issues:
- Prevent memory hungry apps from forcing other apps (eg. sshd) into deep swap. AFAIK setrlimit does that on a per-process level. We need a per-class or per-process-hierarchy level.
- Prevent disk i/o intensive apps from blocking other apps that only do limited disk i/o. Again, we need some class or hierarchy structure to collect the statistics reliably. Please note: AFAIK current disk io schedulers are only "disk input schedulers".
With the release of the kernel.org 2.6.25 kernel, there is now the cgroup extension available. It has already been integrated with the CPU scheduler, and it could also be possible to use it for memory and/or disk i/o rate distribution between, for example, users.